Rewrite Feature_extraction with parallel processing and per-channel alignment

- Parallelize movie processing using joblib (n_jobs=32) for faster extraction
- Fix dual-camera alignment: apply segmentation mask alignment only when the
  raw fluorescence channel is on a different camera than the brightfield.
  Camera 1 (brightfield, 638nm) channels share the mask coordinate space;
  Camera 2 (488nm, 561nm) channels require the inverse calibration transform.
  Previously, the same alignment was applied uniformly to all channels,
  causing misalignment for channels on the same camera as the mask.
- Extract intensity from Channel 2 and Channel 3 when available, with
  per-channel alignment based on wavelength-to-camera mapping from manifest
- Add retry logic with exponential backoff for BioImage loading and dask
  array computation to handle transient network errors
- Support fixed-cell (immunostaining) experiments with correct timepoint
  calculation from fixation time metadata
- Add local file loading option to io.load_imaging_and_segmentation_dataset()
- Add setuptools package discovery config to pyproject.toml
- Add joblib dependency to pyproject.toml
- Improve README with pipeline overview, step descriptions, gene metric
  definitions, and dual-camera alignment explanation

Author: smishra3